无线电工程

针对基于深度学习的端到端调制识别方法存在识别率较低、神经网络参数量大的问题，提出了一种基于时间卷积网络(Temporal Convolutional Network, TCN)的调制识别算法。利用一维因果膨胀卷积，提取信号的时域和频域特征，并加入批归一化(Batch Normalization, BN)和Dropout提高算法的拟合能力；使用全局平均池化(Global Average Pooling, GAP)层替代Flatten层，整合向量特征信息，进一步实现信号的准确分类。在RML 2016.10a数据集上验证了TCN算法的识别性能，在不同信噪比(Signal to Noise Ratio, SNR)下对11种调制信号的平均识别率达到62.3%,与CNN,LSTM和SCRNN算法相比分别提高了10.5%、4.1%、1.1%,参数量分别降低了98.8%、82.5%、91.2%。所提方法对于通信调制信号识别的研究领域具有理论参考价值，对复杂环境下的空间信号智能分类的研究具有工程借鉴意义。

Considering the problem of low recognition rate and large number of parameters in neural networks in end-to-end modulation recognition methods based on deep learning, a TCN modulation recognition algorithm based on Temporal Convolutional Network(TCN)is proposed. This algorithm uses one-dimensional dilated causal convolution to extract the time-domain and frequency-domain features of signals, and adds Batch Normalization(BN) and Dropout to improve the fitting ability of the algorithm; Global Average Pooling(GAP) floor is used to replace the Flatten floor to integrate feature information of vectors and further achieve the accurate classification of signals. The RML 2016.10a dataset is used to validate the recognition performance of TCN algorithm. The algorithm achieves an average recognition rate of 62.3% for 11 modulated signals at various Signal-to-Noise Ratios(Signal to Noise Ratio, SNR),which realizes an improvement of 10.5%,4.1% and 1.1% respectively compared with CNN,LSTM,and SCRNN algorithms, as well as a reduction of parameters by 98.8%,82.5% and 91.2% respectively. The proposed method has theoretical reference value for the research of communication modulation signal recognition, as well as engineering reference significance for the study of intelligent classification of spatial signals in complex environment.